[R-package] confusing error when using lgb.train() with an unknown metric

How you are using LightGBM?

LightGBM component: R package

Environment info

Operating System: macOS 10.14

C++ compiler version: gcc 8.1.0

CMake version: 3.17.3

R version: 4.0.2

LightGBM version or commit hash: https://github.com/microsoft/LightGBM/tree/c07644d1d71540204a9b56f26667e8180bd009e2

Reproducible example(s)

Thanks to @Laurae2 for sharing this with me and creating the reproducible example below.

After installing with Rscript build_r.R, code that uses an unrecognized metric, like this:

library(lightgbm)
library(data.table)
set.seed(1)
labels <- sample(2, 100, replace = TRUE) - 1
data <- as.matrix(data.frame(A = runif(100, min = 0, max = 1), B = runif(100, min = 0, max = 1)))
data_train <- data[1:90, ]
data_valid <- data[91:100, ]
labels_train <- labels[1:90]
labels_valid <- labels[91:100]
dtrain_lgb <- lgb.Dataset(data_train, label = labels_train)
dvalid_lgb <- lgb.Dataset.create.valid(dtrain_lgb, data_valid, label = labels_valid)
valids_lgb <- list(valid = dvalid_lgb)

model <- lgb.train(
    obj = "binary",
    params = list(metric = "nonsense"),
    data = dtrain_lgb,
    valids = valids_lgb,
    nrounds = 1,
    verbose = 1,
    num_thread = 1
)

produces this error

[LightGBM] [Info] Number of positive: 47, number of negative: 43 [LightGBM] [Warning] Auto-choosing col-wise multi-threading, the overhead of testing was 0.000016 seconds. You can set force_col_wise=true to remove the overhead. [LightGBM] [Info] Total Bins 62 [LightGBM] [Info] Number of data points in the train set: 90, number of used features: 2 [LightGBM] [Info] [binary:BoostFromScore]: pavg=0.522222 -> initscore=0.088947 [LightGBM] [Info] Start training from score 0.088947 [LightGBM] [Warning] No further splits with positive gain, best gain: -inf Error in env$eval_list[[1L]] : subscript out of bound

How to close this issue

I think the ideal fix is to raise an error from the C++ side for unrecognized metrics, so that all wrappers benefit from the fix. That would mean changing https://github.com/microsoft/LightGBM/blob/c07644d1d71540204a9b56f26667e8180bd009e2/src/metric/metric.cpp#L64 to raise an error instead of returning a null pointer.

If this change isn't made on the C++ side, I would add a new .METRIC_ALIASES in https://github.com/microsoft/LightGBM/blob/c07644d1d71540204a9b56f26667e8180bd009e2/R-package/R/aliases.R, which lists all of the valid metrics from https://lightgbm.readthedocs.io/en/latest/Parameters.html#metric-parameters, and then raise an error in lgb.check.eval() when any unknown metrics are provided in params.

@Laurae2 @guolinke @StrikerRUS @btrotta what do you think?